Genetic Algorithms Applied to Discrete Distribution Fitting

A common problem when dealing with preprocessing of real world data for a large variety of applications, such as classification and outliers detection, consists in fitting a probability distribution to a set of observations. Traditional approaches often require the resolution of complex equations systems or the use of specialized software for numerical resolution. This paper proposes an approach to discrete distribution fitting based on Genetic Algorithms which is easy to use and has a large variety of potential applications. This algorithm is able not only to identify the discrete distribution function type but also to simultaneously find the optimal value of its parameters. The proposed approach has been applied to an industrial problem concerning surface quality monitoring in flat steel products. The results of the tests, which have been developed using real world data coming from three different industries, demonstrate the effectiveness of the method.